BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.bham.ac.uk//v3//EN
BEGIN:VEVENT
CATEGORIES:Astrophysics Talks Series
SUMMARY:Artificial Neural Networks and Data Analysis - Bru
nello Tirozzi
DTSTART:20191016T130000Z
DTEND:20191016T140000Z
UID:TALK3789AT
URL:/talk/index/3789
DESCRIPTION:I will speak about pattern recognition using Artif
icial Neural Networks (ANN) and SOM (SOM) ( self .
organizing maps). First I define the\ninput and ou
tput patterns and the search of the dimension of t
he patterns. This requires to look for the correla
tion length of the time series..\nI will define th
e training\, validation and learning error. The di
mension of inputs patterns is defined by the cor
relation length.\nThe architecture is defined also
by the number of input neurons\, layers and synap
tic weights.\nThe best architecture is found minim
izing the learning and generalization error studyi
ng the\nvariation of the dependence of these quant
ities on the possible choices a heuristic research
.\nI will show examples of networks with exact est
imates of the learning\nerror. The minimization of
the error can be performed \,both theoretically a
nd experimentally only if the number of patterns\n
is large. The case in which is possible to show r
igourosly the theoretical result holds only for an
infinite number of patterns and infinite number o
f neurons.
LOCATION:PW-SR1 (103)
CONTACT:Dr Matteo Bianconi
END:VEVENT
END:VCALENDAR